# ============================================================================= # vLLM dual-model (GLM-OCR + Qwen3-VL) + PP-DocLayoutV3 (CPU) + Nginx # Hugging Face Space - Nvidia T4-small (16 Go VRAM, sm_75) # Base: official vLLM nightly image, pinned by linux/amd64 digest (2026-11-02). # Verified contents of this digest: vllm 0.33.0rc1.dev484, transformers 5.22.1 # (includes GlmOcrForConditionalGeneration), huggingface-hub 2.4.0, tokenizers 0.23.3. # GLM-OCR needs transformers >= 5.x (arch `glm_ocr`) — stable vLLM images pin # transformers<5 and crash with "model type glm_ocr ... not recognized"; the nightly # ships the whole compatible stack, so NO pip override is needed. # Qwen3-VL only needs vLLM >= 0.11.0 -> covered. sm_75 (T4) wheels included. # Orchestration: supervisord (vllm-ocr, layout, vllm-vlm, warmup, nginx) # ============================================================================= # /! # Digest of the linux/amd64 sub-manifest ONLY. The manifest-list digest # (sha256:d0a6498c...) is the arm64 build -> "exec format error" on amd64 builders. FROM vllm/vllm-openai@sha256:7a933b7cf81799936262bd459ba9a13b589291f9edd6b0179dd7632969906a81 ENV DEBIAN_FRONTEND=noninteractive \ PYTHONUNBUFFERED=1 # SECURITY: non-root user (UID/GID 1000 for HF Spaces compatibility). # The vLLM nightly base image already ships a user/group with UID/GID 1000 # -> create only if missing, and run as numeric UID 1000 (name may differ). RUN getent group 1000 >/dev/null 2>&1 || groupadd -g 1000 space RUN getent passwd 1000 >/dev/null 2>&1 || useradd -u 1000 -g 1000 -m -d /home/space space # System packages: Nginx (reverse proxy). supervisord is NOT taken from apt: # the Ubuntu package (4.2.1) imports asynchat, removed from the stdlib in Python 3.12 # (the base vllm image ships Python 3.12) -> it crashes. We install supervisor via pip below. RUN apt-get update && apt-get install -y --no-install-recommends \ nginx \ curl \ openssl \ ca-certificates \ pciutils \ python3-pip \ && apt-get clean \ && rm -rf /var/lib/apt/lists/* # Python deps for auth + orchestrator + layout service (torch/vllm/huggingface_hub already # in base image). supervisor>=4.2.5 supports Python 3.12 (no asynchat). # PP-DocLayoutV3 runs via ONNX Runtime on CPU: PaddlePaddle is intentionally NOT installed # (it would conflict with torch/vLLM). RUN pip3 install --no-cache-dir \ supervisor \ bcrypt==4.1.2 \ onnxruntime \ pillow \ numpy \ fastapi \ uvicorn \ httpx \ pymupdf \ || pip3 install --no-cache-dir --break-system-packages \ supervisor \ bcrypt==4.1.2 \ onnxruntime \ pillow \ numpy \ fastapi \ uvicorn \ httpx \ pymupdf # Application layout RUN mkdir -p /home/space/.cache \ /tmp/hf \ /data/hf \ /scripts \ /var/log/nginx \ /var/lib/nginx \ /tmp/nginx/body \ /tmp/nginx/proxy \ /tmp/nginx/fastcgi \ /tmp/nginx/uwsgi \ /tmp/nginx/scgi \ /tmp/torch_extensions \ /tmp/.cache # Nginx config, orchestration scripts and supervisord config COPY nginx/ /etc/nginx/ COPY scripts/ /scripts/ COPY supervisord.conf /etc/supervisord.conf RUN chmod +x /scripts/entrypoint.sh /scripts/warmup.sh /scripts/serve_vlm.sh /scripts/serve_nginx.sh /scripts/preload_models.sh /scripts/gpu_detect.py /scripts/generate_htpasswd.py /scripts/layout_service.py && \ chown -R 1000:1000 /home/space /data /scripts /etc/nginx /var/log/nginx /var/lib/nginx /tmp/nginx /tmp/.cache /tmp/hf && \ chmod 777 /tmp/hf && \ chmod -R 755 /etc/nginx /var/log/nginx /var/lib/nginx /tmp/nginx && \ chmod 777 /data && \ chmod 1777 /tmp # Default environment (all overridable via HF Space secrets / docker-compose) ENV HOME=/home/space \ HF_HOME=/tmp/hf \ XDG_CACHE_HOME=/tmp/.cache \ TORCH_EXTENSIONS_DIR=/tmp/torch_extensions \ OMP_NUM_THREADS=1 \ NGINX_PORT=7860 \ MODEL_OCR=zai-org/GLM-OCR \ MODEL_VLM=Qwen/Qwen3-VL-4B-Instruct \ SERVED_OCR=glm-ocr \ SERVED_VLM=qwen3-vl \ PORT_OCR=8001 \ PORT_VLM=8002 \ LAYOUT_PORT=8003 \ LAYOUT_MODEL=PaddlePaddle/PP-DocLayoutV3_onnx \ GPU_MEM_UTIL_OCR=0.25 \ GPU_MEM_UTIL_VLM=0.70 \ MAX_MODEL_LEN=4096 \ MAX_NUM_SEQS=2 \ MAX_IMAGE_PX=2250000 \ MIN_VRAM_MB=15000 \ MIN_CUDA_MAJOR=11 \ MIN_CUDA_MINOR=8 # Hugging Face Space port EXPOSE 7860 # Long start period: nginx now binds port 7860 only after the models are loaded # in VRAM (scripts/serve_nginx.sh), so this check can legitimately fail for a # long time at boot. HEALTHCHECK --interval=30s --timeout=10s --start-period=3600s --retries=3 \ CMD curl -f -u "${NGINX_USERNAME:-admin}:${NGINX_PASSWORD}" http://localhost:7860/health || exit 1 # SECURITY: run as non-root (numeric UID: base image may name the user differently) USER 1000 ENTRYPOINT ["/scripts/entrypoint.sh"]